Novel Moving-Target Detection Using A Hybrid of RGB Images and LiDAR Point-Clouds

Moving objects, especially humans and vehicles, impose very high risks on driverless cars or autonomous driving. Real-time moving-target detection plays a critical role in the future intelligent transportation systems due to public-safety concerns. Recently object tracking and scene analysis using a hybrid of high-resolution RGB images and low-resolution LiDAR point-clouds have been drawing a lot of research interest because fast but precise information rendering can be facilitated thereby. Compared to the conventional object detectors using RGB images, the new target detectors using LiDAR point-clouds can lead to three-dimensional object localization with the crucial depth information RGB images cannot provide, which is very critical to navigation of autonomous vehicles and robots. In this paper, we explore a novel moving-target detection approach which is built upon motion-estimation using RGB images and deep-learning over LiDAR data so that the moving object(s) can be accurately localized in an image frame and the target recognition can be conducted using LiDAR point-clouds. Our proposed new scheme can benefit from both tracking-accuracy brought by RGB images and sufficient training-data together with low time-complexity brought by LiDAR data. Preliminary experimental results have shown that our proposed scheme can automatically identify and track moving human-objects in the scene successively.

运动物体，尤其是人和车辆，给无人驾驶汽车或自动驾驶带来了极高的风险。出于公共安全考虑，实时运动目标检测在未来的智能交通系统中起着至关重要的作用。最近，利用高分辨率RGB图像和低分辨率激光雷达点云混合的物体跟踪和场景分析引起了大量的研究兴趣，因为这样可以促进快速而精确的信息呈现。与使用RGB图像的传统物体探测器相比，使用激光雷达点云的新型目标探测器能够实现具有RGB图像无法提供的关键深度信息的三维物体定位，这对自动驾驶车辆和机器人的导航至关重要。在本文中，我们探索了一种新颖的运动目标检测方法，该方法基于使用RGB图像进行运动估计以及对激光雷达数据进行深度学习，以便能够在图像帧中准确定位运动物体，并利用激光雷达点云进行目标识别。我们提出的新方案可以受益于RGB图像带来的跟踪精度以及激光雷达数据带来的充足训练数据和低时间复杂度。初步实验结果表明，我们提出的方案能够连续自动识别和跟踪场景中的运动人体目标。